Breaking

NVIDIA AI Spotlights H Company Dynamo Stack for Computer-Use Agents

October 8, 2026 at 07:04 EDT

AI Agents
Infra & Chips
Foundation Models

NVIDIA AI is putting H Company's Holo and Holotron computer-use models at the center of an Oct. 8 technical session on how to serve vision-language agents with NVIDIA Dynamo. The focus is not another benchmark demo, but the infrastructure problem behind agents that repeatedly look at a screen, decide what to do, act, and then look again.

Oct. 8, 2026 · NVIDIA AI and H Company

Serving the vision loop

Holo and Holotron computer-use agents put repeated screen perception, action and cache reuse at the center of inference. Dynamo is being positioned as the serving layer for those long-running workloads.

85.2%

Holo4 27B on OSWorld

262K

Token context window

97%

Dynamo cache hit rate ceiling

2-3x

More concurrent sessions

Cost per OSWorld task

$0.05

Holo4 35B-A3B

$0.08

Holo4 27B

Each block represents one cent. The larger 27B model carries higher task cost while posting the stronger OSWorld score.

What improves

Open models, API access and cache-aware routing make long agent sessions easier to inspect, tune and run economically.

What remains hard

Latency, permissions, execution isolation and auditable action logs still determine whether computer control is production-ready.

Agent-serving loop

See screen

→

Choose action

→

Act in app

→

Reuse cache

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.